Sean Silva
> > Those are just hyperparameters. Hyperparameters do not change during program execution and are expected to require different traced/scripted program captures per set of hyperparameters. > > yes I...
@powderluv is there any further action item here or set of hyperparameters we need to cover before calling DLRM "done"?
@powderluv anything else to do here?
I don't think we should special case static shapes. We should do like we do for elementwise broadcasting ops and emit error checks for the "size-1" cases.
Was this ever resolved?
Got it, closing.
Yes, that is where the lowering should go.
No, we aren't adding any custom attributes to ops -- it is already hard enough to manage the zoo of ops and their exact semantics. Adding sideband attributes to the...
I'm still trying to understand better the workflow and details of the use case. Let's iterate in https://github.com/llvm/torch-mlir/issues/910 to get a better understanding of the needs here and then come...
Closing this, as the issues have been moved to other issues or will come up as part of separate quantization efforts.